Update clGetDeviceInfo to follow OpenCL 1.2 spec

Author: sophimao

src/acl_device.cpp

@@ -2,6 +2,7 @@
 // SPDX-License-Identifier: BSD-3-Clause
 
 // System headers.
+#include <algorithm>
 #include <assert.h>
 #include <string.h>
 
@@ -209,6 +210,8 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     RESULT_INT(0);
     break;
   case CL_DEVICE_GLOBAL_MEM_SIZE: {
+#ifdef __arm__
+    // TODO: legacy code here, need to verify correctness with ARM board
     auto gmem_id = acl_get_default_device_global_memory(device->def);
     if (gmem_id < 0) {
       RESULT_INT(0);
@@ -217,10 +220,20 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     cl_ulong size =
         ACL_RANGE_SIZE(device->def.autodiscovery_def.global_mem_defs[gmem_id]
                            .get_usable_range());
-#ifdef __arm__
     // on SoC board, two DDR systems are not equivalent
     // so only half can be accessed with a single alloc.
     size /= 2;
+#else
+    cl_ulong size = 0;
+    for (unsigned gmem_idx = 0;
+         gmem_idx < device->def.autodiscovery_def.num_global_mem_systems;
+         gmem_idx++) {
+      if (device->def.autodiscovery_def.global_mem_defs[gmem_idx].type ==
+          ACL_GLOBAL_MEM_DEVICE_PRIVATE) {
+        size += ACL_RANGE_SIZE(
+            device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
+      }
+    }
 #endif
     RESULT_ULONG(size);
     break;
@@ -251,13 +264,9 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     RESULT_UINT(acl_platform.max_constant_args);
     break;
 
-  // "desktop" profile says global memory must be at least 128MB
-  // "embedded" profile says global memory must be at least 1MB
   case CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE: {
-    // Constant memory is global memory.
-    // However conformance_test_api min_max_constant_buffer_size
-    // expects to allocate two buffers of the size we say here.
-    // So be a shade conservative and cut it down by 4.
+#ifdef __arm__
+    // TODO: legacy code here, need to verify correctness with ARM board
     auto gmem_id = acl_get_default_device_global_memory(device->def);
     if (gmem_id < 0) {
       RESULT_INT(0);
@@ -267,13 +276,32 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
         ACL_RANGE_SIZE(device->def.autodiscovery_def.global_mem_defs[gmem_id]
                            .get_usable_range()) /
         4;
-#ifdef __arm__
-    // see above
+    // Cut by 2 again, see comment for CL_DEVICE_GLOBAL_MEM_SIZE
     size /= 2;
+#else
+    // Constant memory is global memory.
+    cl_ulong size = 0;
+    for (unsigned gmem_idx = 0;
+         gmem_idx < device->def.autodiscovery_def.num_global_mem_systems;
+         gmem_idx++) {
+      if (device->def.autodiscovery_def.global_mem_defs[gmem_idx].type ==
+          ACL_GLOBAL_MEM_DEVICE_PRIVATE) {
+        size += ACL_RANGE_SIZE(
+            device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
+      }
+    }
+    // Note that OpenCL 1.2 specifies the minimum value for devices not of
+    // type CL_DEVICE_TYPE_CUSTOM to be 64KB, however, the OpenCL conformance
+    // test api:test_min_max_constant_buffer_size expects to allocate two
+    // buffers of the size returned here, so following the spec may result
+    // in conformance test failures.
+    size = std::max(size, (cl_ulong)64 * 1024);
 #endif
     RESULT_ULONG(size);
   } break;
   case CL_DEVICE_MAX_MEM_ALLOC_SIZE: {
+#ifdef __arm__
+    // TODO: legacy code here, need to verify correctness with ARM board
     auto gmem_id = acl_get_default_device_global_memory(device->def);
     if (gmem_id < 0) {
       RESULT_INT(0);
@@ -282,7 +310,6 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     cl_ulong size =
         ACL_RANGE_SIZE(device->def.autodiscovery_def.global_mem_defs[gmem_id]
                            .get_usable_range());
-#ifdef __arm__
     // on SoC board, two DDR systems are not equivalent
     // so only half can be accessed with a single alloc.
 
@@ -294,6 +321,28 @@ CL_API_ENTRY cl_int CL_API_CALL clGetDeviceInfoIntelFPGA(
     } else {
       size = size / 8;
     }
+#else
+    cl_ulong global_mem_size = 0;
+    cl_ulong size = 0;
+    for (unsigned gmem_idx = 0;
+         gmem_idx < device->def.autodiscovery_def.num_global_mem_systems;
+         gmem_idx++) {
+      if (device->def.autodiscovery_def.global_mem_defs[gmem_idx].type ==
+          ACL_GLOBAL_MEM_DEVICE_PRIVATE) {
+        global_mem_size += ACL_RANGE_SIZE(
+            device->def.autodiscovery_def.global_mem_defs[gmem_idx].range);
+        cl_ulong curr_size = ACL_RANGE_SIZE(
+            device->def.autodiscovery_def.global_mem_defs[gmem_idx]
+                .get_usable_range());
+        if (curr_size > size) {
+          size = curr_size;
+        }
+      }
+    }
+    // OpenCL 1.2: min value = max(CL_DEVICE_GLOBAL_MEM_SIZE/4, 1*1024*1024)
+    // for devices that are not of type CL_DEVICE_TYPE_CUSTOM
+    size = std::max(size,
+                    std::max(global_mem_size / 4, (cl_ulong)1 * 1024 * 1024));
 #endif
     RESULT_ULONG(size);
   } break;

src/acl_mem.cpp

@@ -5663,7 +5663,7 @@ void acl_mem_migrate_buffer(void *user_data, acl_device_op_t *op) {
         memory_migration.src_mem_list[index].destination_mem_id;
 
 #ifdef MEM_DEBUG_MSG
-    printf("object %d, mem %zx, count %d:", index, (size_t)src_mem,
+    printf("object %d, mem %zx, count %d:\n", index, (size_t)src_mem,
            src_mem->reserved_allocations_count[dest_device][dest_mem_id]);
 #endif